An Empirical Evaluation on Statistical Parsing of Japanese Sentences Using Lexical Association Statistics

An Empirical Evaluation on Statistical Parsing of Japanese Sentences using Lexical Association Statistics SHIRAI Kiyoaki INUI Kentaro TOKUNAGA Takenobu TANAKA Hozumi Departrnent of Computer Science) Graduate School of Information Science and Enginc~t~ring 1 Tokyo Institute of Technology Abstract \Ve are proposing a new framework of statisti- cal language modeling which integrates lexical .association statistics with syntactic preference, while maintaining the modularity of those differ- ent statistics types, facilitating both training of the model and analysis of its behavior. In this paper, we report the result of an empirical evalu- ation of our model, where the model is applied to disambiguation of dependency structures of Japanese sentences. We also discussed the room remained for further improvement based on our error analysis. 1 Introduction In the statistical parsing literature) it has alread:y been established that statistics of lexical associ- ation have real potential for improvement of dis- ambiguation performance. The question is how lexical association statistics should be incorpo- rated into the overall statistical parsing frame- work. In exploring this issue) we consider the following four basic requirements: o Integration of difjeTent types of statistics: Lexical association statistics should be inte- gra.ted with other types of statistics that are also expected to be effective in statistical pars- ing1 such as short-term POS n-gnun statistics and long-term structural preferences over parse trees. • Modularity of statistics types: The total score of a parse derivation should be decomposable into factors derived from differ- ent. types of statistics) which would facilitate analysis of a modePs behavior in terms of each statistics type. • Pmbabilistically well-fottnded semantics: The language model used in a statistical parser should have probabilistically well-founclecl se- mantics) whieh \vould a.lso facilitate the anal:,'·· sis of the model's behavior. , 80 o Trainability: Since incorporation of lexical association statis- tics would make the model prohibitively com- plex, the model's complexity should be flexibly controllable depending on the amount of avail- able training data. However) it seems to be the case that no existing framework of language modeling [2, 4, 12, 1:3, H. 17) 18] satisfies these basic requirements simulta·· neouslyl. In this context 1 we newly designed a framework of statistical language modeling tak- ing all of the above four requirements int.o ac- count [8) 9]. This paper reports on tlw n'sult.:) of our preliminary experiment where our f'rrmH'- vmrk was applied to structural disambigu<Hion of Japanese sentences. In what follows) we first briefly review our framework (Section 2). \Ve next describe tlw sc:t- ting of our experiment) including a brief intro- duction of Japanese dependency struetures, t.hC' data sets1 the baseline of the perfonnanc.C', Nc. (Section 3). We then describe the results of the experinH:mt) which was designed to assess the lu1 .. pact of the the incorporation of lexical associ- ation statistics (Section 4). \Vc finall~· discuss the current problems revealed through our <T- ror analysis, suggesting some possible solutions (Section 5). 2 Overview of our framework As with the most statistical parsing frameworks. given an input string A) we rank …

Content truncated to 500 words